| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| In the Linux kernel, the following vulnerability has been resolved:
net, bpf: fix null-ptr-deref in xdp_master_redirect() for down master
syzkaller reported a kernel panic in bond_rr_gen_slave_id() reached via
xdp_master_redirect(). Full decoded trace:
https://syzkaller.appspot.com/bug?extid=80e046b8da2820b6ba73
bond_rr_gen_slave_id() dereferences bond->rr_tx_counter, a per-CPU
counter that bonding only allocates in bond_open() when the mode is
round-robin. If the bond device was never brought up, rr_tx_counter
stays NULL.
The XDP redirect path can still reach that code on a bond that was
never opened: bpf_master_redirect_enabled_key is a global static key,
so as soon as any bond device has native XDP attached, the
XDP_TX -> xdp_master_redirect() interception is enabled for every
slave system-wide. The path xdp_master_redirect() ->
bond_xdp_get_xmit_slave() -> bond_xdp_xmit_roundrobin_slave_get() ->
bond_rr_gen_slave_id() then runs against a bond that has no
rr_tx_counter and crashes.
Fix this in the generic xdp_master_redirect() by refusing to call into
the master's ->ndo_xdp_get_xmit_slave() when the master device is not
up. IFF_UP is only set after ->ndo_open() has successfully returned,
so this reliably excludes masters whose XDP state has not been fully
initialized. Drop the frame with XDP_ABORTED so the exception is
visible via trace_xdp_exception() rather than silently falling through.
This is not specific to bonding: any current or future master that
defers XDP state allocation to ->ndo_open() is protected. |
| In the Linux kernel, the following vulnerability has been resolved:
dm cache policy smq: fix missing locks in invalidating cache blocks
In passthrough mode, the policy invalidate_mapping operation is called
simultaneously from multiple workers, thus it should be protected by a
lock. Otherwise, we might end up with data races on the allocated blocks
counter, or even use-after-free issues with internal data structures
when doing concurrent writes.
Note that the existing FIXME in smq_invalidate_mapping() doesn't affect
passthrough mode since migration tasks don't exist there, but would need
attention if supporting fast device shrinking via suspend/resume without
target reloading.
Reproduce steps:
1. Create a cache device consisting of 1024 cache entries
dmsetup create cmeta --table "0 8192 linear /dev/sdc 0"
dmsetup create cdata --table "0 131072 linear /dev/sdc 8192"
dmsetup create corig --table "0 262144 linear /dev/sdc 262144"
dd if=/dev/zero of=/dev/mapper/cmeta bs=4k count=1 oflag=direct
dmsetup create cache --table "0 262144 cache /dev/mapper/cmeta \
/dev/mapper/cdata /dev/mapper/corig 128 2 metadata2 writethrough smq 0"
2. Populate the cache, and record the number of cached blocks
fio --name=populate --filename=/dev/mapper/cache --rw=randwrite --bs=4k \
--size=64m --direct=1
nr_cached=$(dmsetup status cache | awk '{split($7, a, "/"); print a[1]}')
3. Reload the cache into passthrough mode
dmsetup suspend cache
dmsetup reload cache --table "0 262144 cache /dev/mapper/cmeta \
/dev/mapper/cdata /dev/mapper/corig 128 2 metadata2 passthrough smq 0"
dmsetup resume cache
4. Write to the passthrough cache. By setting multiple jobs with I/O
size equal to the cache block size, cache blocks are invalidated
concurrently from different workers.
fio --filename=/dev/mapper/cache --name=test --rw=randwrite --bs=64k \
--direct=1 --numjobs=2 --randrepeat=0 --size=64m
5. Check if demoted matches cached block count. These numbers should
match but may differ due to the data race.
nr_demoted=$(dmsetup status cache | awk '{print $12}')
echo "$nr_cached, $nr_demoted" |
| In the Linux kernel, the following vulnerability has been resolved:
bpf, sockmap: Take state lock for af_unix iter
When a BPF iterator program updates a sockmap, there is a race condition in
unix_stream_bpf_update_proto() where the `peer` pointer can become stale[1]
during a state transition TCP_ESTABLISHED -> TCP_CLOSE.
CPU0 bpf CPU1 close
-------- ----------
// unix_stream_bpf_update_proto()
sk_pair = unix_peer(sk)
if (unlikely(!sk_pair))
return -EINVAL;
// unix_release_sock()
skpair = unix_peer(sk);
unix_peer(sk) = NULL;
sock_put(skpair)
sock_hold(sk_pair) // UaF
More practically, this fix guarantees that the iterator program is
consistently provided with a unix socket that remains stable during
iterator execution.
[1]:
BUG: KASAN: slab-use-after-free in unix_stream_bpf_update_proto+0x155/0x490
Write of size 4 at addr ffff8881178c9a00 by task test_progs/2231
Call Trace:
dump_stack_lvl+0x5d/0x80
print_report+0x170/0x4f3
kasan_report+0xe4/0x1c0
kasan_check_range+0x125/0x200
unix_stream_bpf_update_proto+0x155/0x490
sock_map_link+0x71c/0xec0
sock_map_update_common+0xbc/0x600
sock_map_update_elem+0x19a/0x1f0
bpf_prog_bbbf56096cdd4f01_selective_dump_unix+0x20c/0x217
bpf_iter_run_prog+0x21e/0xae0
bpf_iter_unix_seq_show+0x1e0/0x2a0
bpf_seq_read+0x42c/0x10d0
vfs_read+0x171/0xb20
ksys_read+0xff/0x200
do_syscall_64+0xf7/0x5e0
entry_SYSCALL_64_after_hwframe+0x76/0x7e
Allocated by task 2236:
kasan_save_stack+0x30/0x50
kasan_save_track+0x14/0x30
__kasan_slab_alloc+0x63/0x80
kmem_cache_alloc_noprof+0x1d5/0x680
sk_prot_alloc+0x59/0x210
sk_alloc+0x34/0x470
unix_create1+0x86/0x8a0
unix_stream_connect+0x318/0x15b0
__sys_connect+0xfd/0x130
__x64_sys_connect+0x72/0xd0
do_syscall_64+0xf7/0x5e0
entry_SYSCALL_64_after_hwframe+0x76/0x7e
Freed by task 2236:
kasan_save_stack+0x30/0x50
kasan_save_track+0x14/0x30
kasan_save_free_info+0x3b/0x70
__kasan_slab_free+0x47/0x70
kmem_cache_free+0x11c/0x590
__sk_destruct+0x432/0x6e0
unix_release_sock+0x9b3/0xf60
unix_release+0x8a/0xf0
__sock_release+0xb0/0x270
sock_close+0x18/0x20
__fput+0x36e/0xac0
fput_close_sync+0xe5/0x1a0
__x64_sys_close+0x7d/0xd0
do_syscall_64+0xf7/0x5e0
entry_SYSCALL_64_after_hwframe+0x76/0x7e |
| In the Linux kernel, the following vulnerability has been resolved:
bpf: Validate node_id in arena_alloc_pages()
arena_alloc_pages() accepts a plain int node_id and forwards it through
the entire allocation chain without any bounds checking.
Validate node_id before passing it down the allocation chain in
arena_alloc_pages(). |
| In the Linux kernel, the following vulnerability has been resolved:
crypto: ccp - copy IV using skcipher ivsize
AF_ALG rfc3686-ctr-aes-ccp requests pass an 8-byte IV to the driver.
ccp_aes_complete() restores AES_BLOCK_SIZE bytes into the caller's IV
buffer while RFC3686 skciphers expose an 8-byte IV, so the restore
overruns the provided buffer.
Use crypto_skcipher_ivsize() to copy only the algorithm's IV length. |
| In the Linux kernel, the following vulnerability has been resolved:
net/sched: taprio: fix use-after-free in advance_sched() on schedule switch
In advance_sched(), when should_change_schedules() returns true,
switch_schedules() is called to promote the admin schedule to oper.
switch_schedules() queues the old oper schedule for RCU freeing via
call_rcu(), but 'next' still points into an entry of the old oper
schedule. The subsequent 'next->end_time = end_time' and
rcu_assign_pointer(q->current_entry, next) are use-after-free.
Fix this by selecting 'next' from the new oper schedule immediately
after switch_schedules(), and using its pre-calculated end_time.
setup_first_end_time() sets the first entry's end_time to
base_time + interval when the schedule is installed, so the value
is already correct.
The deleted 'end_time = sched_base_time(admin)' assignment was also
harmful independently: it would overwrite the new first entry's
pre-calculated end_time with just base_time. |
| In the Linux kernel, the following vulnerability has been resolved:
pppoe: drop PFC frames
RFC 2516 Section 7 states that Protocol Field Compression (PFC) is NOT
RECOMMENDED for PPPoE. In practice, pppd does not support negotiating
PFC for PPPoE sessions, and the current PPPoE driver assumes an
uncompressed (2-byte) protocol field. However, the generic PPP layer
function ppp_input() is not aware of the negotiation result, and still
accepts PFC frames.
If a peer with a broken implementation or an attacker sends a frame with
a compressed (1-byte) protocol field, the subsequent PPP payload is
shifted by one byte. This causes the network header to be 4-byte
misaligned, which may trigger unaligned access exceptions on some
architectures.
To reduce the attack surface, drop PPPoE PFC frames. Introduce
ppp_skb_is_compressed_proto() helper function to be used in both
ppp_generic.c and pppoe.c to avoid open-coding. |
| In the Linux kernel, the following vulnerability has been resolved:
netfilter: nf_tables: join hook list via splice_list_rcu() in commit phase
Publish new hooks in the list into the basechain/flowtable using
splice_list_rcu() to ensure netlink dump list traversal via rcu is safe
while concurrent ruleset update is going on. |
| In the Linux kernel, the following vulnerability has been resolved:
bonding: 3ad: implement proper RCU rules for port->aggregator
syzbot found a data-race in bond_3ad_get_active_agg_info /
bond_3ad_state_machine_handler [1] which hints at lack of proper
RCU implementation.
Add __rcu qualifier to port->aggregator, and add proper RCU API.
[1]
BUG: KCSAN: data-race in bond_3ad_get_active_agg_info / bond_3ad_state_machine_handler
write to 0xffff88813cf5c4b0 of 8 bytes by task 36 on cpu 0:
ad_port_selection_logic drivers/net/bonding/bond_3ad.c:1659 [inline]
bond_3ad_state_machine_handler+0x9d5/0x2d60 drivers/net/bonding/bond_3ad.c:2569
process_one_work kernel/workqueue.c:3302 [inline]
process_scheduled_works+0x4f0/0x9c0 kernel/workqueue.c:3385
worker_thread+0x58a/0x780 kernel/workqueue.c:3466
kthread+0x22a/0x280 kernel/kthread.c:436
ret_from_fork+0x146/0x330 arch/x86/kernel/process.c:158
ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:245
read to 0xffff88813cf5c4b0 of 8 bytes by task 22063 on cpu 1:
__bond_3ad_get_active_agg_info drivers/net/bonding/bond_3ad.c:2858 [inline]
bond_3ad_get_active_agg_info+0x8c/0x230 drivers/net/bonding/bond_3ad.c:2881
bond_fill_info+0xe0f/0x10f0 drivers/net/bonding/bond_netlink.c:853
rtnl_link_info_fill net/core/rtnetlink.c:906 [inline]
rtnl_link_fill+0x1d7/0x4e0 net/core/rtnetlink.c:927
rtnl_fill_ifinfo+0xf8e/0x1380 net/core/rtnetlink.c:2168
rtmsg_ifinfo_build_skb+0x11c/0x1b0 net/core/rtnetlink.c:4453
rtmsg_ifinfo_event net/core/rtnetlink.c:4486 [inline]
rtmsg_ifinfo+0x6d/0x110 net/core/rtnetlink.c:4495
__dev_notify_flags+0x76/0x390 net/core/dev.c:9790
netif_change_flags+0xac/0xd0 net/core/dev.c:9823
do_setlink+0x905/0x2950 net/core/rtnetlink.c:3180
rtnl_group_changelink net/core/rtnetlink.c:3813 [inline]
__rtnl_newlink net/core/rtnetlink.c:3981 [inline]
rtnl_newlink+0xf55/0x1400 net/core/rtnetlink.c:4109
rtnetlink_rcv_msg+0x64b/0x720 net/core/rtnetlink.c:6995
netlink_rcv_skb+0x123/0x220 net/netlink/af_netlink.c:2550
rtnetlink_rcv+0x1c/0x30 net/core/rtnetlink.c:7022
netlink_unicast_kernel net/netlink/af_netlink.c:1318 [inline]
netlink_unicast+0x5a8/0x680 net/netlink/af_netlink.c:1344
netlink_sendmsg+0x5c8/0x6f0 net/netlink/af_netlink.c:1894
sock_sendmsg_nosec net/socket.c:787 [inline]
__sock_sendmsg net/socket.c:802 [inline]
____sys_sendmsg+0x563/0x5b0 net/socket.c:2698
___sys_sendmsg+0x195/0x1e0 net/socket.c:2752
__sys_sendmsg net/socket.c:2784 [inline]
__do_sys_sendmsg net/socket.c:2789 [inline]
__se_sys_sendmsg net/socket.c:2787 [inline]
__x64_sys_sendmsg+0xd4/0x160 net/socket.c:2787
x64_sys_call+0x194c/0x3020 arch/x86/include/generated/asm/syscalls_64.h:47
do_syscall_x64 arch/x86/entry/syscall_64.c:63 [inline]
do_syscall_64+0x12c/0x3b0 arch/x86/entry/syscall_64.c:94
entry_SYSCALL_64_after_hwframe+0x77/0x7f
value changed: 0x0000000000000000 -> 0xffff88813cf5c400
Reported by Kernel Concurrency Sanitizer on:
CPU: 1 UID: 0 PID: 22063 Comm: syz.0.31122 Tainted: G W syzkaller #0 PREEMPT(full)
Tainted: [W]=WARN
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 04/18/2026 |
| In the Linux kernel, the following vulnerability has been resolved:
net: tls: fix strparser anchor skb leak on offload RX setup failure
When tls_set_device_offload_rx() fails at tls_dev_add(), the error path
calls tls_sw_free_resources_rx() to clean up the SW context that was
initialized by tls_set_sw_offload(). This function calls
tls_sw_release_resources_rx() (which stops the strparser via
tls_strp_stop()) and tls_sw_free_ctx_rx() (which kfrees the context),
but never frees the anchor skb that was allocated by alloc_skb(0) in
tls_strp_init().
Note that tls_sw_free_resources_rx() is exclusively used for this
"failed to start offload" code path, there's no other caller.
The leak did not exist before commit 84c61fe1a75b ("tls: rx: do not use
the standard strparser"), because the standard strparser doesn't try
to pre-allocate an skb.
The normal close path in tls_sk_proto_close() handles cleanup by calling
tls_sw_strparser_done() (which calls tls_strp_done()) after dropping
the socket lock, because tls_strp_done() does cancel_work_sync() and
the strparser work handler takes the socket lock. |
| In the Linux kernel, the following vulnerability has been resolved:
net: ena: PHC: Fix potential use-after-free in get_timestamp
Move the phc->active check and resp pointer assignment to after
acquiring the spinlock. Previously, phc->active was checked without
holding the lock, and resp was cached from ena_dev->phc.virt_addr
before the lock was acquired.
If ena_com_phc_destroy() runs between the lockless active check and
the lock acquisition, it sets active=false, releases the lock, frees
the DMA memory, and sets virt_addr=NULL. The get_timestamp path would
then read a NULL virt_addr and dereference it.
With both the active check and the pointer read under the lock,
destroy cannot free the memory while get_timestamp is using it. |
| In the Linux kernel, the following vulnerability has been resolved:
KVM: s390: pci: fix GAIT table indexing due to double-scaling pointer arithmetic
kvm_s390_pci_aif_enable(), kvm_s390_pci_aif_disable(), and
aen_host_forward() index the GAIT by manually multiplying the index
with sizeof(struct zpci_gaite).
Since aift->gait is already a struct zpci_gaite pointer, this
double-scales the offset, accessing element aisb*16 instead of aisb.
This causes out-of-bounds accesses when aisb >= 32 (with
ZPCI_NR_DEVICES=512)
Fix by removing the erroneous sizeof multiplication. |
| In the Linux kernel, the following vulnerability has been resolved:
smb/client: fix possible infinite loop and oob read in symlink_data()
On 32-bit architectures, the infinite loop is as follows:
len = p->ErrorDataLength == 0xfffffff8
u8 *next = p->ErrorContextData + len
next == p
On 32-bit architectures, the out-of-bounds read is as follows:
len = p->ErrorDataLength == 0xfffffff0
u8 *next = p->ErrorContextData + len
next == (u8 *)p - 8 |
| In the Linux kernel, the following vulnerability has been resolved:
ceph: put folios not suitable for writeback
The batch holds references to the folios (see `filemap_get_folios`,
`folio_batch_release`), so we need to `folio_put` the folios we remove.
Tested on v6.18. |
| In the Linux kernel, the following vulnerability has been resolved:
libceph: handle rbtree insertion error in decode_choose_args()
A message of type CEPH_MSG_OSD_MAP contains an OSD map that itself
contains a CRUSH map. The received CRUSH map may optionally contain
choose_args that get decoded in decode_choose_args(). In this function,
num_choose_arg_maps is read from the message, and a corresponding number
of crush_choose_arg_maps gets decoded afterwards. Each
crush_choose_arg_map has a choose_args_index, which serves as the key
when inserting it into the choose_args rbtree of the decoded crush_map.
If a (potentially corrupted) message contains two crush_choose_arg_maps
with the same index, the assertion in insert_choose_arg_map() triggers a
kernel BUG when trying to insert the second crush_choose_arg_map.
This patch fixes the issue by switching to the non-asserting rbtree
insertion function and rejecting the message if the insertion fails.
[ idryomov: changelog ] |
| In the Linux kernel, the following vulnerability has been resolved:
iommu/vt-d: Fix oops due to out of scope access
Below oops triggers when kill QEMU process:
Oops: general protection fault, probably for non-canonical address 0x7fffffff844eaaa7: 0000 [#1] SMP NOPTI
Call Trace:
<TASK>
do_raw_spin_lock+0xaa/0xc0
_raw_spin_lock_irqsave+0x21/0x40
domain_remove_dev_pasid+0x52/0x160
intel_nested_set_dev_pasid+0x1b9/0x1e0
__iommu_set_group_pasid+0x56/0x120
pci_dev_reset_iommu_done+0xe3/0x180
pcie_flr+0x65/0x160
__pci_reset_function_locked+0x5b/0x120
vfio_pci_core_close_device+0x63/0xe0 [vfio_pci_core]
vfio_df_close+0x4f/0xa0
vfio_df_unbind_iommufd+0x2d/0x60
vfio_device_fops_release+0x3e/0x40
__fput+0xe5/0x2c0
task_work_run+0x58/0xa0
do_exit+0x2c8/0x600
do_group_exit+0x2f/0xa0
get_signal+0x863/0x8c0
arch_do_signal_or_restart+0x24/0x100
exit_to_user_mode_loop+0x87/0x380
do_syscall_64+0x2ff/0x11e0
entry_SYSCALL_64_after_hwframe+0x76/0x7e
The global static blocked domain is a dummy domain without corresponding
dmar_domain structure, accessing beyond iommu_domain structure triggers
oops easily. Fix it by return early in domain_remove_dev_pasid() like
identity domain. |
| In the Linux kernel, the following vulnerability has been resolved:
drm/xe/dma-buf: handle empty bo and UAF races
There look to be some nasty races here when triggering the
invalidate_mappings hook:
1) We do xe_bo_alloc() followed by the attach, before the actual full bo
init step in xe_dma_buf_init_obj(). However the bo is visible on the
attachments list after the attach. This is bad since exporter driver,
say amdgpu, can at any time call back into our invalidate_mappings hook,
with an empty/bogus bo, leading to potential bugs/crashes.
2) Similar to 1) but here we get a UAF, when the invalidate_mappings
hook is triggered. For example, we get as far as xe_bo_init_locked()
but this fails in some way. But here the bo will be freed on error, but
we still have it attached from dma-buf pov, so if the
invalidate_mappings is now triggered then the bo we access is gone and
we trigger UAF and more bugs/crashes.
To fix this, move the attach step until after we actually have a fully
set up buffer object. Note that the bo is not published to userspace
until later, so not sure what the comment "Don't publish the bo
until we have a valid attachment", is referring to.
We have at least two different customers reporting hitting a NULL ptr
deref in evict_flags when importing something from amdgpu, followed by
triggering the evict flow. Hit rate is also pretty low, which would
hint at some kind of race, so something like 1) or 2) might explain
this.
v2:
- Shuffle the order of the ops slightly (no functional change)
- Improve the comment to better explain the ordering (Matt B)
(cherry picked from commit af1f2ad0c59fe4e2f924c526f66e968289d77971) |
| In the Linux kernel, the following vulnerability has been resolved:
net: qrtr: fix refcount saturation and potential UAF in qrtr_port_remove
In qrtr_port_remove(), the socket reference count is decremented via
__sock_put() before the port is removed from the qrtr_ports XArray and
before the RCU grace period elapses.
This breaks the fundamental RCU update paradigm. It exposes a race
window where a concurrent RCU reader (such as qrtr_reset_ports() or
qrtr_port_lookup()) can obtain a pointer to the socket from the XArray,
and attempt to call sock_hold() on a socket whose reference count has
already dropped to zero.
This exact race condition was hit during syzkaller fuzzing, leading to
the following refcount saturation warning and a potential Use-After-Free:
refcount_t: saturated; leaking memory.
WARNING: CPU: 3 PID: 1273 at lib/refcount.c:22 refcount_warn_saturate+0xae/0x1d0
Modules linked in: qrtr(+) bochs drm_shmem_helper ...
Call Trace:
<TASK>
qrtr_reset_ports net/qrtr/af_qrtr.c:768 [inline] [qrtr]
__qrtr_bind.isra.0+0x48b/0x570 net/qrtr/af_qrtr.c:805 [qrtr]
qrtr_bind+0x17d/0x210 net/qrtr/af_qrtr.c:901 [qrtr]
kernel_bind+0xe4/0x120 net/socket.c:3592
qrtr_ns_init+0x1a6/0x380 net/qrtr/ns.c:715 [qrtr]
qrtr_proto_init+0x3b/0xff0 net/qrtr/af_qrtr.c:169 [qrtr]
do_one_initcall+0xf5/0x5e0 init/main.c:1283
...
</TASK>
Fix this by deferring the reference count decrement until after the
xa_erase() and the synchronize_rcu() complete.
(Note: The v1 of this patch incorrectly replaced __sock_put() with
sock_put(). As Simon Horman pointed out, the callers of qrtr_port_remove()
still hold a reference to the socket, so freeing the socket memory here
would lead to a subsequent UAF in the caller. Thus, the __sock_put() is
kept, but only repositioned to close the RCU race.) |
| In the Linux kernel, the following vulnerability has been resolved:
fs/fcntl: fix SOFTIRQ-unsafe lock order in fasync signaling
A SOFTIRQ-safe to SOFTIRQ-unsafe lock order deadlock can occur in
send_sigio() and send_sigurg() when a process group receives a signal.
When FASYNC is configured for a process group (PIDTYPE_PGID), both
functions use read_lock(&tasklist_lock) to traverse the task list.
However, they are frequently called from softirq context:
- send_sigio() via input_inject_event -> kill_fasync
- send_sigurg() via tcp_check_urg -> sk_send_sigurg (NET_RX_SOFTIRQ)
The deadlock is caused by the rwlock writer fairness mechanism:
1. CPU 0 (process context) holds read_lock(&tasklist_lock) in do_wait().
2. CPU 1 (process context) attempts write_lock(&tasklist_lock) in
fork() or exit() and spins, which blocks all new readers.
3. CPU 0 is interrupted by a softirq (e.g., TCP URG packet reception).
4. The softirq calls send_sigurg() and attempts to acquire
read_lock(&tasklist_lock), deadlocking because CPU 1 is waiting.
Since PID hashing and do_each_pid_task() traversals are already
RCU-protected, the read_lock on tasklist_lock is no longer strictly
required for safe traversal. Fix this by replacing tasklist_lock with
rcu_read_lock(), aligning the process group signaling path with the
single-PID path. This also mitigates a potential remote denial of
service vector via TCP URG packets.
Lockdep splat:
=====================================================
WARNING: SOFTIRQ-safe -> SOFTIRQ-unsafe lock order detected
[...]
Chain exists of:
&dev->event_lock --> &f_owner->lock --> tasklist_lock
Possible interrupt unsafe locking scenario:
CPU0 CPU1
---- ----
lock(tasklist_lock);
local_irq_disable();
lock(&dev->event_lock);
lock(&f_owner->lock);
<Interrupt>
lock(&dev->event_lock);
*** DEADLOCK *** |
| In the Linux kernel, the following vulnerability has been resolved:
Revert "wireguard: device: enable threaded NAPI"
This reverts commit 933466fc50a8e4eb167acbd0d8ec96a078462e9c which is
commit db9ae3b6b43c79b1ba87eea849fd65efa05b4b2e upstream.
We have had three independent production user reports in combination
with Cilium utilizing WireGuard as encryption underneath that k8s Pod
E/W traffic to certain peer nodes fully stalled. The situation appears
as follows:
- Occurs very rarely but at random times under heavy networking load.
- Once the issue triggers the decryption side stops working completely
for that WireGuard peer, other peers keep working fine. The stall
happens also for newly initiated connections towards that particular
WireGuard peer.
- Only the decryption side is affected, never the encryption side.
- Once it triggers, it never recovers and remains in this state,
the CPU/mem on that node looks normal, no leak, busy loop or crash.
- bpftrace on the affected system shows that wg_prev_queue_enqueue
fails, thus the MAX_QUEUED_PACKETS (1024 skbs!) for the peer's
rx_queue is reached.
- Also, bpftrace shows that wg_packet_rx_poll for that peer is never
called again after reaching this state for that peer. For other
peers wg_packet_rx_poll does get called normally.
- Commit db9ae3b ("wireguard: device: enable threaded NAPI")
switched WireGuard to threaded NAPI by default. The default has
not been changed for triggering the issue, neither did CPU
hotplugging occur (i.e. 5bd8de2 ("wireguard: queueing: always
return valid online CPU in wg_cpumask_choose_online()")).
- The issue has been observed with stable kernels of v5.15 as well as
v6.1. It was reported to us that v5.10 stable is working fine, and
no report on v6.6 stable either (somewhat related discussion in [0]
though).
- In the WireGuard driver the only material difference between v5.10
stable and v5.15 stable is the switch to threaded NAPI by default.
[0] https://lore.kernel.org/netdev/CA+wXwBTT74RErDGAnj98PqS=wvdh8eM1pi4q6tTdExtjnokKqA@mail.gmail.com/
Breakdown of the problem:
1) skbs arriving for decryption are enqueued to the peer->rx_queue in
wg_packet_consume_data via wg_queue_enqueue_per_device_and_peer.
2) The latter only moves the skb into the MPSC peer queue if it does
not surpass MAX_QUEUED_PACKETS (1024) which is kept track in an
atomic counter via wg_prev_queue_enqueue.
3) In case enqueueing was successful, the skb is also queued up
in the device queue, round-robin picks a next online CPU, and
schedules the decryption worker.
4) The wg_packet_decrypt_worker, once scheduled, picks these up
from the queue, decrypts the packets and once done calls into
wg_queue_enqueue_per_peer_rx.
5) The latter updates the state to PACKET_STATE_CRYPTED on success
and calls napi_schedule on the per peer->napi instance.
6) NAPI then polls via wg_packet_rx_poll. wg_prev_queue_peek checks
on the peer->rx_queue. It will wg_prev_queue_dequeue if the
queue->peeked skb was not cached yet, or just return the latter
otherwise. (wg_prev_queue_drop_peeked later clears the cache.)
7) From an ordering perspective, the peer->rx_queue has skbs in order
while the device queue with the per-CPU worker threads from a
global ordering PoV can finish the decryption and signal the skb
PACKET_STATE_CRYPTED out of order.
8) A situation can be observed that the first packet coming in will
be stuck waiting for the decryption worker to be scheduled for
a longer time when the system is under pressure.
9) While this is the case, the other CPUs in the meantime finish
decryption and call into napi_schedule.
10) Now in wg_packet_rx_poll it picks up the first in-order skb
from the peer->rx_queue and sees that its state is still
PACKET_STATE_UNCRYPTED. The NAPI poll routine then exits e
---truncated--- |
